Question

我想在Ubuntu 16.04上构建TensorFlow 1.3（而不是1.13），并支持MPI（而不是默认的gRPC）。我从Ubuntu仓库安装了软件包libopenmpi-dev。运行/usr/lib/openmpi脚本时，我已经提供了configure作为MPI工具包目录。

我使用此命令启动构建：

$ bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package

但是存在标头包含问题：

文件tensorflow/contrib/mpi/mpi_utils.cc包含tensorflow/contrib/mpi/mpi_utils.h
此mpi_utils.h包括third_party/mpi/mpi.h
此mpi.h是指向/usr/lib/openmpi/include/mpi.h的符号链接
实际的mpi.h包含以下行：

#include "openmpi/ompi/mpi/cxx/mpicxx.h"

并且mpicxx.h位于文件夹/usr/lib/openmpi/include/openmpi/ompi/mpi/cxx/中，该文件夹不在包含路径中。

我已通过创建指向正确文件夹的符号链接来“修复”此问题：

$ ln -s /usr/lib/openmpi/include/openmpi third_party/mpi/openmpi

现在找到了mpicxx.h，但是它想包含mpi.h，由于mpi.h不在同一个文件夹中，因此失败了：

$ bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package
WARNING: /home/arno/tensorflow/tensorflow/contrib/learn/BUILD:15:1: in py_library rule //tensorflow/contrib/learn:learn: target '//tensorflow/contrib/learn:learn' depends on deprecated target '//tensorflow/contrib/session_bundle:exporter': No longer supported. Switch to SavedModel immediately.
WARNING: /home/arno/tensorflow/tensorflow/contrib/learn/BUILD:15:1: in py_library rule //tensorflow/contrib/learn:learn: target '//tensorflow/contrib/learn:learn' depends on deprecated target '//tensorflow/contrib/session_bundle:gc': No longer supported. Switch to SavedModel immediately.
INFO: Found 1 target...
ERROR: /home/arno/tensorflow/tensorflow/contrib/mpi/BUILD:60:1: C++ compilation of rule '//tensorflow/contrib/mpi:mpi_rendezvous_mgr' failed: crosstool_wrapper_driver_is_not_gcc failed: error executing command external/local_config_cuda/crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc -U_FORTIFY_SOURCE '-D_FORTIFY_SOURCE=1' -fstack-protector -fPIE -Wall -Wunused-but-set-parameter ... (remaining 151 argument(s) skipped): com.google.devtools.build.lib.shell.BadExitStatusException: Process exited with status 1.
In file included from ./third_party/mpi/mpi.h:2673:0,
                 from ./tensorflow/contrib/mpi/mpi_utils.h:27,
                 from ./tensorflow/contrib/mpi/mpi_rendezvous_mgr.h:33,
                 from tensorflow/contrib/mpi/mpi_rendezvous_mgr.cc:18:
./third_party/mpi/openmpi/ompi/mpi/cxx/mpicxx.h:35:17: fatal error: mpi.h: No such file or directory
compilation terminated.
Target //tensorflow/tools/pip_package:build_pip_package failed to build
Use --verbose_failures to see the command lines of failed build steps.
ERROR: /home/arno/tensorflow/tensorflow/tools/pip_package/BUILD:134:1 C++ compilation of rule '//tensorflow/contrib/mpi:mpi_rendezvous_mgr' failed: crosstool_wrapper_driver_is_not_gcc failed: error executing command external/local_config_cuda/crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc -U_FORTIFY_SOURCE '-D_FORTIFY_SOURCE=1' -fstack-protector -fPIE -Wall -Wunused-but-set-parameter ... (remaining 151 argument(s) skipped): com.google.devtools.build.lib.shell.BadExitStatusException: Process exited with status 1.
INFO: Elapsed time: 6.668s, Critical Path: 4.98s

我尝试使用以下命令将标头的路径手动添加到GCC的包含路径：

$ bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package --copt='-I/usr/lib/openmpi/include'

...但是随后出现错误，因为/usr/lib/openmpi/include/openmpi/ompi/mpi/cxx中包含的标头未在Bazel的配置文件中声明。而且我无法将它们声明给Bazel，因为它不接受绝对路径。

我找不到使此构建工作正常的方法。我是Bazel的新手，根据我的阅读，构建目录应该是“独立的”，即包含所有必需的头文件和源文件，但是TensorFlow存储库通过向{{1}添加符号链接来违反此规定。 /usr/lib/...中的}。不能更改TensorFlow版本。

如何通过OpenMPI支持构建TensorFlow 1.3？

编辑：将third_party/mpi选项添加到Bazel build命令中，如注释中所建议的那样，将提供更详细的输出，但是我无法确定使用哪个编译器。我认为这些是相关的行：

-s

Answer 1

一种解决方法是从源代码构建和安装MVAPICH（然后，MPI工具箱路径为/usr/local）。该问题仅存在于OpenMPI中。

构建TensorFlow：从Bazel根外部导入MPI标头

1 个答案: